Diagnostic screening of urban soil contaminants using diffuse reflectance spectroscopy.
Soil contamination demands efficient methods for diagnosis and remediation. This is largely due to the potential health risks to humans and ecosystems. Comprehensive assessments of contaminated land must include the identification of contaminants and a risk assessment of exposure. Public awareness has resulted in many government and public authorities finding themselves under increased pressure to ensure that any discovery of contaminated soils is conveyed and remediated quickly and effectively. Therefore, efficient techniques for sampling and analyses are necessary (Tiller 1992; Chuang et al. 2003). Due to the heterogeneity of some soils, there is great difficulty in identifying both significant areas of contamination and their sources (Hawley 1985).
The Australian and New Zealand Environment and Conservation Council (ANZECC) determined soil contamination thresholds, called environmental investigation 'trigger values'. In most cases they differ from the National Health and Medical Research Council (NHMRC) values. The ANZECC trigger values for lead, zinc, copper, cadmium, benzo[a]pyrene (BaP), and total polycyclic aromatic hydrocarbons (PAH) are 300, 200, 120, 3, 1, and 20mg/kg, respectively (ANZECC 1992). Total PAH implies the total concentration of the 16 PAH that have been identified as particularly carcinogenic or mutagenic, and includes BaP (Xing et al. 2006). The health effects of the heavy metals mentioned in concentrations above the ANZECC investigation levels include kidney damage (cadmium and copper), inhibited brain development (lead), and anaemia (lead and zinc) (Imray and Langley 1999).
The contaminants discussed above present challenges to those looking to efficiently diagnose, analyse, and rehabilitate contaminated sites. That their toxicity can vary so rapidly over a very short space in landscapes such as urban or industrial areas is also challenging (Markus 2000). Since urban and industrial land uses now compete in many urban areas, such as in Sydney's inner suburbs, continued local use of urban soil has become even more important. Approaches to the spatial modelling of heavy metal contamination have mostly involved various forms of geostatistics (e.g. Goovaerts 1999) and spatial point pattern analysis (Walter et al. 2005).
Despite their high level of accuracy, many conventional methods of contaminant identification are costly and time-consuming (e.g. soil PAH identification, Jensen et al. 2007). In light of the demand for rapid and cheap contaminant identification there is a niche for diffuse reflectance spectroscopy (DRS) to be introduced into identification of land contaminants (Kooistra et al. 2007). The technology is simple to use and poses no health threats to the user, unlike other spectroscopic technologies that require protective equipment to be worn (e.g. X-ray spectroscopy). DRS is non-destructive and a single scan, combined with some multivariate statistical analysis, can measure several properties simultaneously (Viscarra Rossel et al. 2006). Perhaps most importantly, soil analysis can be completed within minutes and the analytical cost is low. The aim of this paper is to develop a methodology for the diagnostic screening of heavy metals and PAH using vis-NIR and MIR DRS.
Two soil datasets were used for the experiment. The first set contained 489 samples from the Sydney inner-west suburb of Glebe, New South Wales, with known concentrations of cadmium, copper, lead, and zinc obtained from studies by Markus and McBratney (1996). The second set contained 65 samples with known PAH concentrations from an inner city suburb in Melbourne, Victoria. Concentrations of heavy metals and PAH were determined by atomic absorption spectrometry (Rayment and Higginson 1992).
Diffuse reflectance spectroscopy
Diffuse reflectance spectra of the heavy metal dataset were recorded in the vis-NIR range, from 350 to 2500 nm, and in the MIR range, from 2500 to 25 000nm (or 4000 to 400[cm.sup.-l]). The spectrometers used were the vis-NIR Agrispec (ASD Inc.) and the FTIR Tensor 37 (Bruker Corporation). Spectra were collected from the vis-NIR range at 1-nm intervals averaging 10 scans collected per second and from the MIR range at 8 cm I and 64 scans per minute. A Spectralon white reference was used in calibration for the vis-NIR spectrometer, while spectroscopic grade potassium bromide (KBr) was used for MIR spectrometer calibrations.
Soil samples were ground to <2 mm for the vis-NIR analysis and to <200 [micro]m for the MIR analysis. The spectra of the PAH dataset were measured using the vis-NIR spectrometer only, as these samples were moist and volatile in nature, making MIR analysis using the Tensor 37 and our current setup more difficult.
The analysis was carried out for the vis-N1R and mid-IR separately. The spectra were transformed to log 1/reflectance, and a principal component analysis (PCA) was used to compress the spectra into fewer principal components. The first 10 principal component scores were retained and a selection of these was used as the independent variables in the development of the regressions, described below.
For each of the contaminants the data were coded into binary 0 or 1, describing uncontaminated or contaminated sites, respectively, at a particular threshold. The selection of thresholds was based on current ANZECC guidelines: Zn 200mg/kg, Pb 300mg/kg, Cu 60mg/kg, Cd 3mg/kg, total PAH 20mg/kg, and BaP 1 mg/kg. We also coded data for values equivalent to half and double the ANZECC thresholds. For example, the binary coding of Pb at the current ANZECC threshold of 300 mg/kg was conducted as follows: any samples containing lead in concentrations [greater than or equal to] 300mg/kg were coded as '1' and any samples with concentrations below this threshold were coded '0'. Therefore, for Pb we also coded data for half and double the ANZECC threshold, corresponding to 150 and 600mg/kg, respectively.
Ordinal logistic regressions (OLR) were used for the screening of soil contamination (Hastie et al. 2001). We used this model because we wanted to model the posterior probabilities of the binary classes (0 uncontaminated and 1 contaminated), using linear functions of the spectra, while simultaneously ensuring that they sum to 1 and remain in [0, 1]. The principal component scores were used to represent the spectra in the models and were thus the independent variables used to predict the presence or absence of contamination at a particular threshold. Selection of which, and how many, of these principal components to use was made using a stepwise selection procedure that used a forward selection and backward elimination technique using a probability of <0.1 to accept or discard a predictor variable (i.e. a principal component).
Each dataset was randomly split into 2 independent subsets: 2/3 for 'training' and the remaining for 'testing' the models. The OLR models were derived using the 'training' data and the models were independently tested using the remaining 1/3 'testing' dataset. Thus, the OLR models were used to predict whether or not the soils were contaminated based on the predetermined thresholds.
Diagnostic screening of soil contaminants
Type I errors, or false positives, and Type II errors, or false negatives were determined using contingency tables (Fig. 1). Type I errors occur when a sample has been predicted to be contaminated when it is in fact uncontaminated (Allchin 2001). This has cost implications with respect to remediation or further monitoring processes. Soil that is falsely identified as contaminated will be treated as a health risk and may undergo expensive remediation or ongoing monitoring. The proportion of uncontaminated sites reported as contaminated, the false positive rate (FPR), can be calculated from:
FPR = number of false positives/ number of negative instances
Type II errors occur when contaminated samples are predicted to be free of contamination (Allchin 2001). This can have serious consequences, as soils that should be remediated or carefully monitored have neither restriction placed upon them and pose a potential health risk if the contaminant is present in sufficient concentrations. The proportion of contaminated sites reported as uncontaminated, the false negative rate (FNR), can be calculated from:
FNR = number of false negatives/ number of positive instances
Thresholds may be varied to make the test either more restrictive or more sensitive, depending on the design of the experiment. In the development of diagnostic screening tests, it is clearly desirable to lower both the FNR and FPR as much as possible.
Receiver operating characteristic (ROC) curves measure the efficiency of the model's fitted probabilities to sort the response levels. ROC curves were used to assess how the choice of threshold affects the FPR and FNR (Hastie et al. 2001). A ROC curve is a plot of sensitivity by (1 specificity) for each condition. The sensitivity is the probability of correctly predicting contamination when a site is contaminated; and the specificity, is the probability of correctly predicting an uncontaminated site when the site is uncontaminated. Then, the probability of incorrectly predicting contamination is 1--specificity. The area under the ROC curve is a common index used to summarise and assess the fitted model - the higher the ROC curve from the diagonal, the better the model fit.
Exploratory data analysis
Figure 2 shows highly skewed data for all contaminants, which is an intrinsic feature of soil contamination (Markus and McBratney 1996). It would be alarming, for example, if a distribution of a contaminant did not favour the uncontaminated or very low contamination end of the distribution, particularly in an area as large as the 220-ha site examined for the heavy metal study. Lead and the total PAH data, in particular, are highly skewed, with maximum values 70 and 100 times the threshold, respectively.
Both PAH distributions depict >65% of the data above the relevant thresholds; >50% of the samples exhibit copper, lead, and zinc contamination; and 2.5% of samples are contaminated with cadmium. Areas with lead >25 times the ANZECC threshold have been discussed by Markus (2000), and were a result of direct dumping of lead soldering material into the soil. Heavy metal analysis by Markus and McBratney (1996) also showed a strong correlation between contaminations across different samples; where lead contamination was present, there was an increased chance of copper and zinc contamination, and this is also attributed to the previous industrial soil use in Glebe (Markus 2000). However, the distributions show that very few, if any, samples with cadmium or zinc thresholds beyond the NHRMC thresholds. Further analysis was therefore carried out using the ANZECC thresholds only (Fig. 2).
[FIGURE 2 OMITTED]
[FIGURE 3 OMITTED]
The mean absorbance of lead in the vis-NIR range at 3 soil concentrations is shown in Fig. 3a, high (>500mg/kg), low (<150mg/kg), and moderate (150-500mg/kg). The samples that fall into these categories are plotted onto a scores plot for PCA in Fig. 3b. Together, these 2 PCs account for 90.2% of the data variance. The variance plots also show that spectral differentiation is possible between highly contaminated soil and soils that are relatively contaminant-free. This is particularly clear in Fig. 3b, where there is segregation between the 2 groups.
Although Fig. 3a shows very little spectral activity between 350 and 700 nm, clear peaks, characteristic of hydroxyl groups (1430nm) and water and clay (1930nm and 2250nm), do appear. Figure 3a also reveals a difference in absorbance between samples with low and high contamination, reflecting findings by Kemper and Sommer (2002). In low concentrations, heavy metals and PAH are not spectrally active in the visible or infrared portions of the electromagnetic spectrum (Wu et al. 2007) but they can be identified in contaminated soils through the correlations and interactions between the contaminants and spectrally active soil components such as iron, clay, and organic carbon. In this study, the heavy metal contaminants were well correlated with soil organic carbon (Table 1) and there was some correlation between copper and clay content.
[FIGURE 4 OMITTED]
The spectral analysis of total PAH shows a large and prominent peak at 1930nm due to the water in samples containing PAH (Fig. 4a). The sparse distribution of the total PAH data across the principal components scores plot (Fig. 4b) suggests that more samples may be needed for more rigorous examination. The lead clusters shown in Fig. 4a were also fitted to MIR data in Fig. 5. From Figs 3-5, absorbance generally increased with increased soil contamination.
Diagnostic screening using ordinal logistic regression
The contingency table of OLR results for lead contamination in both the vis-NIR and MIR spectra at 3 different thresholds shows that as thresholds increase the true negatives increase while the true positives decrease (Fig. 6). From this observation, it follows that the proportion of false positives decreases while the proportion of false negatives increases with an increase in threshold. This is to be expected given that at lower thresholds, there is an increased likelihood of a sample being contaminated, while the reverse is expected at higher thresholds.
[FIGURE 5 OMITTED]
[FIGURE 6 OMITTED]
The same trend of false negatives and false positives is also observed for zinc (Fig. 7). The change in these values with threshold is more dramatic, with a very small proportion (<4%) of false negatives at lower thresholds (Fig. 7a, d) but a higher proportion of false negatives (>16%) at higher thresholds (Fig. 7c, f). Figure 8 shows contingency tables for both the vis-NIR and MIR analysis of copper.
The OLR results for total PAH from the vis-NIR (Fig. 9) show that no false negatives were predicted for a total PAH threshold at 10 mg/kg, which indicates that we may need more samples for this analysis. Nevertheless, an increase in false negatives with thresholds is shown, but unlike Figs 6 and 7, the proportion of false positives also increases with thresholds. The example of BaP analysis (Fig. 10) shows similar trends for false negatives but is inconclusive for false positive percentage, as an increase is seen from 0.5 to 1 mg/kg but a decrease follows from 1 to 2 mg/kg. This indicates again that more samples are required for PAH analysis.
[FIGURE 7 OMITTED]
[FIGURE 8 OMITTED]
[FIGURE 9 OMITTED]
The model's ability to predict contaminants was measured in 3 ways: the percentage of points correctly predicted (accuracy), the FPR, and the FNR (Table 2). Note that cadmium contamination was not present in sufficient samples to carry out these analyses in full, and PAH samples were not analysed in the MIR range. For both vis-N1R and MIR predictions, there was a decrease in FPR with an increase in threshold and an increase in FNR with an increase in threshold. The predictions of lead were the least accurate of the heavy metals analysed (Table 2), falling to <68% in the vis-NIR analysis at the 300mg/kg threshold. The difference in accuracy between vis-NIR and MIR analyses at the ANZECC threshold and half the ANZECC threshold was noticeable for lead, with 10% difference at 300mg/kg. However, the vis-NIR analysis was slightly more accurate at double the ANZECC threshold of 600mg/kg. The most accurate prediction of heavy metal presence was for zinc at the lowest threshold, with accuracies of 86% and 89.5% in the vis-NIR and MIR, respectively. Copper was the only element to have all 6 threshold predictions exceed 76% accuracy (Table 2). The average accuracy of prediction for copper was 81.63%, exceeding those of zinc (78.9%), lead (75.9%), and total PAH (78.9%). The copper analyses had very high FNR values at the higher threshold of 120mg/kg (0.92). With just 2.5% of the data contaminated with cadmium, predictions for this contaminant were not considered as reliable as other analyses. This was despite the high accuracy of predictions (>90%) in both the vis-NIR and MIR ranges. The FNR was negligible, which could be attributed to the small percentage of samples with cadmium contamination. Previous research has failed to sufficiently quantify or qualify the presence of cadmium in soil (Wu et al. 2005), so it was surprising that this method improved the accuracy of predicting cadmium in the soil. With the exception of lead at a 600mg/kg threshold, all MIR predictions were more accurate than those of vis-N1R, although these differences were as little as 1.3% for zinc at the highest threshold and 1% for lead at the highest threshold. There was little to separate the best threshold at which to predict contamination, although the best predictions were found at the lowest thresholds for all but the copper predictions from the vis-NIR data.
[FIGURE 10 OMITTED]
The range in accuracy of total PAH predictions was the greatest (Table 2). A 25% difference is apparent from half the ANZECC threshold of 10mg/kg to double the threshold at 40mg/kg. The FPR trend was fairly consistent but revealed a rare occurrence of the FPR increasing with the threshold (Table 2), which may be attributed to a lack of samples, given that each threshold bracket had a different number of samples. Accuracy was comparable between PAH and heavy metals at full and half thresholds, although heavy metals were better predicted at the highest threshold. On average, MIR had lower FPR and FNR than did vis-NIR.
Receiver operator curves were used to explore how the choice of a probability threshold affects the proportion of FPR and FNR. ROC curves for each of the contaminants at each threshold (Fig. 11) show that there were small differences between the accuracies of the 3 lead, zinc, and copper thresholds in both NIR and MIR. The strength of the MIR predictions of copper values compares favourably to the NIR predictions (Fig. l le, f) There was also limited success of the PAH predictions, with the 'curves' particularly disjointed as a result of the small sample size (Fig. 1 lg, h). These results suggest that further development could lead to practical applications of DRS to predict soil heavy metal contamination.
Spectroscopic diagnostic screening tests for heavy metal concentration, particularly lead, zinc, and copper, are possible because of their relationship with clay, organic carbon, and iron (Kemper and Sommer 2002). The improved predictions at lower thresholds are encouraging, as it gives a larger margin for error. For example, a false negative from a screening test for lead at a 150mg/kg threshold may not be a contaminated site according to the ANZECC guidelines. Identification of false negatives at the lower thresholds could help to determine which of these false negatives are below the ANZECC guidelines. Cattle et al. (2002) noted that a site where the FNR is >0.85 is probably contaminated but anything <0.45 becomes harder to classify. This may mean that the acceptable levels of false positives and negatives become subjective (Cattle et al. 2002). This idea could be applied to both the vis-NIR and MIR false negative rates of 0.92 and 0.77, respectively. Contaminated site work requires an accuracy of 10-15%. One possible method improvement to our methodology will be to combine the spectroscopic screening tests with some spatial analysis. If a predicted site is contaminated but is surrounded by uncontaminated sites, this may increase the likelihood of the contaminated site being a false positive. Similarly, if a site is predicted to be free of contamination but is surrounded by contaminated sites, then the likelihood that this is a false negative prediction may increase.
While research has indicated a varying degree of effectiveness for screening different contaminants, the advantage of using DRS in the vis-NIR and MIR is that data are analysed in 'real-time', measurements can be made in situ, and several contaminants can be analysed simultaneously. This may also be valuable for analysis of volatile organics, such as PAH, that may diminish in concentration in transitu. Furthermore, the infrared spectrometer contains no dangerous components, which is an advantage over commercial portable X-ray spectrometers, which require protective equipment and training due to low-level X-ray radiation.
Conventional techniques for measuring PAH generally have more than a 5-day turnaround. Each sample may cost upwards of $150, particularly if a faster turnaround is required. Even so-called 'rapid' methods may take more than an hour, as with the identification process outlined by Hua et al. (2007). The spectroscopic process outlined here is clearly different in terms of the cost and time required.
[FIGURE 11 OMITTED]
The use of conventional methods cannot, however, be discarded. In building the model, considerable expense may have to be allocated to determining sample contamination with conventional methods. However, combining a DRS screening method with conventional methods for 'ground-truthing' could prove accurate and far more cost-effective than other conventional methods. This technology will allow for more rapid and cheaper testing of soil contaminants and an increased spatial coverage.
* Despite differences in spectral reflectance, samples with different levels of contamination do not exhibit characteristic absorption features. Spectroscopic predictions are possible because of the interaction of the contaminants with other soil constituents, such as clay, iron, and organic carbon.
* Identification of the level of contamination is possible using principal component analysis.
* Ordinal logistic regression can be used as the method for rapid diagnostic screening of soil contamination. Model accuracies were consistently >75% and up to 90%, in line with the ANZECC thresholds.
* There is a trade-off between the false positive rate (FPR) and false negative rate (FNR), with the former generally decreasing with thresholds and the latter increasing with thresholds. Appropriate guidelines need to be formed regarding acceptable FPR and FNR.
* The MIR range, with an average prediction accuracy of 79.9%, was generally, but not significantly, more accurate for predicting contamination than the vis-NIR range, with an average prediction accuracy of 77.12%.
* The ability of DRS to simultaneously identify several contaminants rapidly, cheaply, and with an increased spatial coverage of sampling should make it appealing to the consultancy sector, provided that some improvements are made in the precision and accuracy of the methods.
We thank Julie Markus and Dahmon Sorongan for the soil samples.
Manuscript received 31 March 2008, accepted 13 February 2009
Allchin D (2001) Error types. Perspectives on Science 9, 38-58. doi: 10.1162/10636140152947786
Australian and New Zealand Environment and Conservation Council (1992) Australian and New Zealand Guidelines for the Assessment and Management of Contaminated Sites. 1992. NHMRC ref. EH17 (rescinded in 2002).
Cattle JA, McBratney AB, Minasny B (2002) Kriging method evaluation for assessing the spatial distribution of urban soil lead contamination. Journal of Environmental Quality 31,1576-1588.
Chuang JC, Van Emon JM, Chou YL, Junod N, Finegold JK, Wilson NK (2003) Comparison of immunoassay and gas chromatography-mass spectrometry for measurement of polycyclic aromatic hydrocarbons in contaminated soil. Analytica Chimica Acta 486, 31-39. doi: 10.1016/ S0003-2670(03)00499-9
Goovaerts P (1999) Geostatistics in soil science: state-of-the-art and perspectives. Geoderma 89, 1-45. doi: 10.1016/S0016-7061(98) 00078-0
Hastie T, Tibshirani R, Friedman J (2001) 'The elements of statistical learning. Data mining, inference, and prediction.' (Springer: New York)
Hawley JK (1985) Assessment of health risk from exposure to contaminated soil. Risk Analysis 5, 289 302. doi: 10.1111/j.1539-6924. 1985.tb00185.x
Hua G, Broderick J, Semple KT, Killham K, Singleton 1 (2007) Rapid quantification of polycyclic aromatic hydrocarbons in hydroxypropyl-[beta]-cyclodextrin (HPCD) soil extracts by synchronous fluorescence spectroscopy (SFS). Environmental Pollution 148, 176-181. doi: 10.1016/j.envpol.2006.10.040
Imray P, Langley A (1999) 'Health-based soil investigation levels.' National Environmental Health Forum monographs. Soil series, no. I. (Department of Health and Ageing: Canberra, ACT)
Jensen H, Reimann C, Finne TE, Ottesen RT, Amoldussen A (2007) PAH-concentrations and compositions in the top 2 cm of forest soils along a 120 km long transect through agricultural areas, forests and the city of Oslo, Norway. Environmental Pollution 145, 829-838. doi: 10.1016/j.envpol.2006.05.008
Kemper T, Sommer S (2002) Estimate of heavy metal contamination in soils after a mining accident using reflectance spectroscopy. Environmental Science & Technology 36, 2742 2747. doi: 10.1021/ es015747j
Kooistra L, Wehrens R, Leuven RSEW, Buydens LMC (2007) Possibilities of visible--near-infrared spectroscopy for the assessment of soil contamination in river floodplains. Analytica Chimica Acta 446, 97 105. doi: 10.1016/S0003-2670(01)01265-X
Markus JA (2000) Study of lead in the soil of Glebe and its environs. PhD Thesis, The University of Sydney, NSW, Australia.
Markus JA, McBratney AB (1996) An urban soil study: heavy metals in Glebe, Australia. Australian Journal of Soil Research 34, 453-465. doi: 10.1071/SR9960453
Rayment GE, Higginson FR (1992) 'Australian laboratory handbook of soil and water chemical methods.' (Inkata Press: Melbourne, Vic.)
Tiller KG (1992) Urban soil contamination in Australia. Australian Journal of Soil Research 30, 937457. doi: 10.1071/SR9920937
Viscarra Rossel RA, Walvoort D J J, McBratney AB, Janik L J, Skjemstad JO (2006) Visible, near infrared, mid infrared or combined diffuse reflectance spectroscopy for simultaneous assessment of various soil properties. Geoderma 131, 59-75. doi: 10.1016/j.geoderma.2005.03.007
Walter C, McBratney AB, Viscarra Rossel RA, Markus JA (2005) Spatial point-process statistics: concepts and application to the analysis of lead contamination in urban soil. Environmetrics 16, 339 355. doi: 10.1002/env.705
Wu Y, Chen J, Ji J, Gong P, Liao Q, Tian Q, Ma H (2007) A mechanism study of reflectance spectroscopy for investigating heavy metals in soils. Soil Science Society of America Journal 71,918 926.
Wu Y, Chen J, Wu X, Tian Q, Ji J, Qin Z (2005) Possibilities of reflectance spectroscopy for the assessment of contaminant elements in suburban soils. Applied Geochemistry 20, 1051-1059. doi: 10.1016/ j.apgeochem.2005.01.009 Xing W, Luo Y, Wu L, Song J, Qian W, Christie P (2006) Spatial distribution of PAHs in a contaminated valley in Southeast China. Environmental Geochemistry and Health 28, 89 96. doi: 10.1007/s10653-005-9017-x
J. G. P. Bray (A), R. Viscarra Rossel (B,C) and A. B. McBratney (A)
(A) Australian Centre for Precision Agriculture, Faculty of Agriculture, Food & Natural Resources, The University of Sydney, NSW 2006, Australia.
(B) CSIRO Land & Water, Bruce E. Butler Laboratory, GPO Box 1666, Canberra, ACT 2601, Australia.
(C) Corresponding author. Email: email@example.com
Table 1. Correlations between heavy metals, clay, and organic carbon Heavy metal units are mg/kg when calculating correlations Analyte % Clay % OC Pb Pb 0.05 0.35 1.00 Cd -0.02 0.46 0.23 Zn 0.01 0.38 0.42 Cu 0.13 0.38 0.09 Analyte Cd Zn Cu Pb 0.23 0.42 0.09 Cd 1.00 0.39 0.49 Zn 0.39 1.00 0.29 Cu 0.49 0.29 1.00 Table 2. Selected contaminants and thresholds used to predict accuracy, false positive rates (FPR), and false negative rates (FNR) from the test set Highest accuracies are highlighted in bold; lowest accuracies are highlighted in italics; ND, lack of contaminated samples with which to run trial vis-NIR predictions Threshold Contaminant (mg/kg) Accuracy (%) FPR Pb 150 76.06 0.53 Pb 300 67.84 0.32 Pb 600 73.07 0.10 Zn 100 86.04 0.66 Zn 200 73.82 0.42 Zn 400 73.07 0.08 Cu 30 83.79 0.49 Cu 60 76.06 0.20 Cu 120 84.29 0.02 Cd 0.5 68.08 0.19 Cd 3.0 ND Total PAH 10 90.25# 0.57 Total PAH 20 80.48 0.88 Total PAH 40 65.85 0.91 BaP 0.5 75.91 0.50 BaP 1 90.25# 0.24 BaP 2 65.90 0.44 vis-NIR average (%) vis-NIR MIR predictions predictions Contaminant FNR Accuracy (%) Pb 0.10 82.65# Pb 0.32 77.17 Pb 0.64 72.60 Zn 0.04 89.5# Zn 0.15 76.71 Zn 0.71 74.42 Cu 0.06 85.39# Cu 0.28 77.70 Cu 0.92 82.65 Cd 0.52 74.88# Cd ND Total PAH 0 Total PAH 0.03 Total PAH 0.13 BaP 0 BaP 0.25 BaP 0.26 77.51 MIR average (%) MIR predictions Contaminant FPR FNR Pb 0.33 0.10 Pb 0.25 0.21 Pb 0.15 0.52 Zn 0.51 0.03 Zn 0.35 0.16 Zn 0.14 0.49 Cu 0.43 0.06 Cu 0.21 0.24 Cu 0.03 0.77 Cd 0.15 0.40 Cd Total PAH Total PAH Total PAH BaP BaP BaP 79.87 Highest accuracies are highlighted in bold indicated with #. Fig. 1. Contingency tables used for diagnostic screening of soil contamination. Test result Actual condition Negative Positive Absent Present Condition absent + Condition present + positive result = false positive result = true positive positive Condition absent + Condition present + negative result = true negative result = false negative negative
|Printer friendly Cite/link Email Feedback|
|Author:||Bray, J.G.P.; Rossel, R. Viscarra; McBratney, A.B.|
|Publication:||Australian Journal of Soil Research|
|Date:||Jul 1, 2009|
|Previous Article:||Mineralogy and geochemical properties of some upland soils from different sedimentary formations in south-eastern Nigeria.|
|Next Article:||Nitrous oxide and methane emissions from soil are reduced following afforestation of pasture lands in three contrasting climatic zones.|